Merge pull request espnet#5657 from wanchichen/patch-1

Add E-Branchformer model for FLEURS

egs2/fleurs/asr1/README.md

@@ -44,6 +44,37 @@ Finally, the model needs to be configured to use the data.
 # RESULTS
 
 <!-- Generated by scripts/utils/show_asr_result.sh -->
+
+# Multilingual ASR - SSL + E-Branchformer + Self-condition [XLS-R, E-Branchformer, utt_mvn, 6500 BPE](conf/tuning/train_asr_ebf_scctc.yaml)
+
+## Environments
+- date: `Tue Feb  7 05:54:39 CST 2023`
+- python version: `3.8.16 (default, Jan 17 2023, 23:13:24)  [GCC 11.2.0]`
+- espnet version: `espnet 202211`
+- pytorch version: `pytorch 1.13.1+cu116`
+- Git hash: `7f37bf7270017eede7d77701b389d1412f30078c`
+  - Commit date: `Sun Jan 1 13:06:01 2023 -0500`
+- Pre-trained model: https://huggingface.co/espnet/wanchichen_fleurs_asr_ebf_scctc
+
+## asr_train_asr_branchformer_scctc_raw_all_bpe6500_sp
+### WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|lm0.4_pen0.0_lm_lm_train_lm_all_bpe6500_valid.loss.ave_asr_model_valid.acc.ave/test_all|77809|1669969|75.2|22.3|2.5|3.0|27.8|95.7|
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|lm0.4_pen0.0_lm_lm_train_lm_all_bpe6500_valid.loss.ave_asr_model_valid.acc.ave/test_all|77809|10235271|93.1|4.3|2.6|2.3|9.2|95.6|
+
+### TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|lm0.4_pen0.0_lm_lm_train_lm_all_bpe6500_valid.loss.ave_asr_model_valid.acc.ave/test_all|77809|9622352|92.2|5.1|2.6|2.4|10.2|95.5|
+
 # Multilingual ASR - SSL + Conformer + Hierarchical LID Self-condition [XLS-R, Conformer, utt_mvn, 6500 BPE](conf/train_asr_conformer_hier_lid_utt.yaml)
 
 ## Environments

egs2/fleurs/asr1/conf/tuning/train_asr_ebf_scctc.yaml

@@ -0,0 +1,94 @@
+batch_type: numel
+batch_bins: 40000000
+accum_grad: 4
+max_epoch: 20
+patience: none
+# The initialization method for model parameters
+init: xavier_uniform
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 3
+
+encoder: e_branchformer
+encoder_conf:
+    output_size: 512
+    attention_heads: 8
+    attention_layer_type: rel_selfattn
+    pos_enc_layer_type: rel_pos
+    rel_pos_type: latest
+    cgmlp_linear_units: 3072
+    cgmlp_conv_kernel: 31
+    use_linear_after_conv: false
+    gate_activation: identity
+    num_blocks: 12
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.1
+    input_layer: conv2d2
+    layer_drop_rate: 0.1
+    linear_units: 1024
+    positionwise_layer_type: linear
+    macaron_ffn: true
+    use_ffn: true
+    merge_conv_kernel: 31
+    interctc_layer_idx: [3, 6, 9]
+    interctc_use_conditioning: true
+
+decoder: transformer
+decoder_conf:
+    attention_heads: 8
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.1
+    src_attention_dropout_rate: 0.1
+
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1
+    interctc_weight: 0.5
+    length_normalized_loss: false
+    extract_feats_in_collect_stats: false
+
+optim: adam
+optim_conf:
+    lr: 0.002
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 25000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
+
+freeze_param: [
+"frontend.upstream"
+]
+
+frontend: s3prl
+frontend_conf:
+    frontend_conf:
+        upstream: wav2vec2_url  # Note: If the upstream is changed, please change the input_size in the preencoder.
+        path_or_url: https://huggingface.co/s3prl/converted_ckpts/resolve/main/xlsr2_300m.pt
+    download_dir: ./hub
+    multilayer_feature: True
+
+preencoder: linear
+preencoder_conf:
+    input_size: 1024  # Note: If the upstream is changed, please change this value accordingly.
+    output_size: 80